1 / 33

Mahesh Balakrishnan, John Davis, Dahlia Malkhi , Vijayan Prabhakaran, Ted Wobber

C lusters O f R aw F lash U nits. Mahesh Balakrishnan, John Davis, Dahlia Malkhi , Vijayan Prabhakaran, Ted Wobber Microsoft Research Silicon Valley Thanks: Phil Bernstein, Ken Eguro, Jeremy Elson , Ed Nightingale, Colin Reid, Michael Wei, Ming Wu. tape is dead disk is tape

panella
Download Presentation

Mahesh Balakrishnan, John Davis, Dahlia Malkhi , Vijayan Prabhakaran, Ted Wobber

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Clusters Of Raw Flash Units Mahesh Balakrishnan, John Davis, Dahlia Malkhi, Vijayan Prabhakaran, Ted Wobber Microsoft Research Silicon Valley Thanks: Phil Bernstein, Ken Eguro, Jeremy Elson , Ed Nightingale, Colin Reid, Michael Wei, Ming Wu

  2. tape is dead disk is tape flash is disk RAM localityis king - Jim Gray, Dec 2006

  3. flash in the data center can flash clusters eliminate the trade-off between consistency and performance? what new abstractions are required to manage and access flash clusters?

  4. what is CORFU? application infrastructure applications: key-value stores databases SMR (stmach rep) filesystems virtual disks application TOFU (trans over) append(value) read(offset) CORFU Replicated, fault-tolerant append read from anywhere append to tail network-attached flash 1 Gbps, 15W, 75 µs reads, 200 µs writes (4KB) cluster of 32 server-attached SSDs, 1.5 TB 0.5M reads/s, 0.2M appends/s (4KB) flash cluster

  5. the case for CORFU applications append/read data why a shared log interface? • easy to build strongly consistent (transactional) in-memory applications • effective way to pool flash: • SSD uses logging to avoid write-in-place • random reads are fast • GC is feasible

  6. CORFU is a distributed SSD with a shared log interface getting 10 Gbps random-IO from 1TB flash farm:

  7. the CORFU design mapping resides at the client CORFU API: read(O) append(V) trim(O) application CORFU append throughput: # of 64-bit tokens issued per second CORFU library read from anywhere append to tail sequencer 4KB log entry each logical entry is mapped to a replica set of physical flash pages

  8. CORFU throughput (server+SSD)[CORFU: A Shared Log Design for Flash Clusters, NSDI 2012] capped at sequencer bottleneck scales linearly

  9. the CORFU protocol: mapping client application read(pos) D1 D3 D5 D7 CORFU library D2 D4 D6 D8 read(D1/D2, page#) Projection: D1 D2 D3 D4 D5 D6 D7 D8 CORFU cluster

  10. the CORFU protocol: tail-finding client CORFU append throughput: # of 64-bit tokens issued per second reserve next position in log (e.g., 100) sequencer (T0) application read(pos) append(val) D1 D3 D5 D7 CORFU library D2 D4 D6 D8 write(D1/D2, val) Projection: D1 D2 D3 D4 D5 D6 D7 D8 CORFU cluster

  11. the CORFU protocol: (chain) replication client C1 client C2 client C3 safety under contention: if multiple clients try to write to same log position concurrently, only one ‘wins’ durability: data is only visible to reads if entire chain has seen it

  12. handling failures: clients • client obtains token from sequencer and crashes:holes in the log • solution: other clients can ‘fill’ the hole • fast fill operation in CORFU API (<1ms): • completes half-written entries • writes junk on unwritten entries (metadata operation, conserves flash cycles, bandwidth)

  13. handling failures: flash units each Projection is a list of views 0 - D1 D2 D3 D4 D5 D6 D7 D8 0 - 7 D1 a D3 D4 D5 D6 D7 D8 8 - D1 D9 D3 D4 D5 D6 D7 D8 0 - 7 D1 a D3 D4 D5 D6 D7 D8 8 – 9 D1 D9 D3 D4 D5 D6 D7 D8 9 - D10 D11 D12 D13 D14 D15 D16 D17 Projection 0 Projection 1 Projection 2 D10 D12 D14 D16 D1 D3 D5 D7 D9 D2 D4 D6 D8 D11 D13 D15 D17

  14. handling failures: sequencer the sequencer is an optimization • clients can instead probe for tail of the log log tail is soft state • value can be reconstructed from flash units • sequencer identity is stored in Projection

  15. from State-Machine-Replication to CORFU[From Paxos to CORFU: A Flash-Speed Shared Log, ACM SIGOPS, 2012] SMR CORFU decide on sequence of log-entry values CORFU-chain: first in chain proposes value (due to seq’cer, no contention) each entry has a designated replica-set throughput capped at rate of sequencing 64-bit tokens scale-out by partitioning over time rather than over data • form a sequence of agreement decisions • leader-based solutions: • leader proposes each entry to group of replicas • same leader, same replicas, for all entries • throughput capped at leader IO capacity • scale-out through partitioning: • multiple autonomous sequences Corfu Paxos

  16. how far is CORFU from Paxos? Paxos instance Paxos instance Paxos instance • CORFU partitions across time, not space • (CORFU replica chain == Paxos instance)

  17. how do we build CORFU applications?

  18. TOFU: lock-free transactions with CORFU application where is your data? TOFU library read/update (key, version, value) did you and I do all reads/updates atomically? versioned table store (BigTable,HBase, PacificA, FDS)

  19. TOFU: lock-free transactions with CORFU application lookup TOFU library read/update (key, version, value) entries versioned table store (BigTable,HBase, PacificA, FDS) update index

  20. TOFU: lock-free transactions with CORFU application lookup TOFU library CORFU library read/update (key, version, value) entries append “commit” intention- record construct index by“playing” log versioned table store (BigTable,HBase, PacificA, FDS)

  21. TOFU: lock-free transactions with CORFU application lookup TOFU library CORFU library read/update (key, version, value) entries append “commit” intention- record construct index by“playing” log versioned table store (BigTable,HBase, PacificA, FDS) ref conflict zone intent

  22. CORFU applications key-value store what: data + commit why: multi-put/get, snapshots, mirroring database (Hyder, Bernstein et al.) what: speculative commit records why: decide commit/abort virtual disk storage what: volume sectors why: multiplex flash cycles / IOPS / bytes state machine replication what: proposed commands why: order/persist commands others: metadata services, distributed STM, filesystems…

  23. grieving for HD • Denial: • too expensive, vulnerable, wait for PCM • Anger: • SSD-FTL, use flash as a block-device • random writes exhibit bursts in performance • duplicate functionality • Bargaining: • use flash as memory-cache/memory-extension • Depression: • throw money and power at it, PCI-e controllers • Acceptance: • new design for farms of flash in a cluster

  24. CORFU Beehive-based flash unit[A Design for Network Flash, NVMW 2012] • Beehive architecture overBEE3 and XUPv5 FPGAs • Multiple slow (soft) cores • Token ring interconnect • Vehicle for practicalarchitectureresearch

  25. Conclusion CORFU is a distributed SSD: 1 million write IOPS, linearly scalable read IOPS CORFU is a shared log: strong consistency at wire speed CORFU uses network-attached flash to construct inexpensive, power-efficient clusters

  26. backup slides

  27. reconfiguration to reconfigure, a client must: • seal the current Projection on drives • determine last log position modified in current Projection • propose a new Projection to the auxiliary as the next Projection an auxiliary stores a sequence of Projections • propose kth Projection; fails if it already exists • read kth Projection; fails if none exists why an auxiliary? • can tolerate f failures with f+1 replicas • can be implemented via conventional Paxos reconfiguration latency = tens of milliseconds

  28. flash unit design • device exposes sparse linear address space • three message types: read, write, seal • write-once semantics: • reads must fail on unwritten pages • writes must fail on already written pages • bad-block remapping • Three form factors: • FPGA+SSD: implemented, single unit • server+SSD: implemented, 32-unit deployment • FPGA+flash: under development

  29. the FPGA+SSD flash unit • BeeHive architecture over BEE3 and XUPv5 FPGAs • BEE3: 80 GB SSD, 8 GB RAM • power: 15W • tput: 800 Mbps • end-to-end latency: • reads: 400 µsecs • mirrored appends: 800 µsecs

  30. garbage collection • once appended, data does not move in the log • application is required to trim invalid positions • log address space can become sparse ∞ invalid entries valid entries

  31. the CORFU cluster (server+SSD) • $3K of flash • 32 drives, 16 servers, 2 racks • Intel X25V (40 GB) • 20,000 4KB read IOPS • each server limited by 1 Gbps NIC at 30,000 4KB IOPS • mirrored across racks max read tput: 500K 4KB IOPS (= 16 Gbps = 2 GB/s) max write tput: 250K 4KB IOPS

  32. flash unit design - message-passing ring of cores - logical separation of functionality + specialized cores - slow cores (100 MHz) = low power (15W)

  33. flash primer • page granularity (e.g., 4KB) • fast random reads • overwriting 4KB page requires surrounding 512KB block to be erased • erasures are slow • erasures cause wear-out (high BER) • SSDs: LBA-over-log • no erasures in critical path • even wear-out across drive

More Related